Method and limits
How an entry is made, and what it means
This page is written for someone who will cite Atlas and wants to know what they are standing on.
1. Sources come first
An operator adds a source document with its title, publisher, publication date and link. Atlas stores a hash of the file, so a later change to the file is detectable. The text of a source is treated as material to read, never as instructions. If a document tries to give the extraction model orders, they are ignored.
2. Events are proposed, and must quote
An AI model reads the text in small pieces and may propose events. Each proposal must carry one exact, unbroken quote from the piece it read. A program checks that the quote really is in the text, word for word. If it is not, the event is thrown away. The model is told to leave things out when unsure, to avoid words that interpret, and never to infer a cause, a motive or a missing date.
3. A person approves what readers see
Proposed events sit in a waiting state. A reviewer approves, rejects or flags each one. Readers only ever see approved events. Approval means “the quote supports this short description”. It does not mean the source is right. Sources can be wrong, biased or disputed, which is why the source is always one click away.
4. Labels on a trajectory
Each trajectory carries one label for how the cited sources say it ended: completed, not completed or open. A curator sets it from the text and writes a short note saying which text supports it. The label describes an endpoint. It is not a judgement of skill, effort or worth, and Atlas uses no red or green for it on purpose.
5. Counts, never odds
The comparison view counts how many events of each type are recorded for each trajectory. It does not turn those counts into shares, rates, scores or probabilities. The set is selected by people, not sampled, so such numbers would mean nothing.
6. What is out of scope
Atlas is a history and evidence product. It does not cover markets, trading, investing or personal finance, and it records no share prices, valuations, returns or forecasts. Business and economic history is welcome as plain historical fact (for example, that an organisation was founded or closed in a stated year). Material that reads as investment insight is stopped before it reaches the model and sent to a person.
7. Known limits
- The library is small and hand-picked. It is not a sample of history, so it cannot show how common anything is.
- Text scanned from old books can carry scanning errors. Atlas keeps the scan as it is and does not silently repair it.
- An AI model helped to propose entries. People approved them. Neither is infallible, which is why every entry links to its source.
- A trajectory that a source calls complete may still be disputed by other sources.